Ated as the coefficient is 0.949 and 0.958 respectively. The coefficients of SVM, RELIEF and LAC with frequency are greater than 0.75 indicating that all are correlated with frequency. Among them, LAC has the strongest correlation (0.947) with frequency. This is mainly caused by bit 3, 8, 11, 36 and 166. For bit 3, 8 and 11, since their frequencies are not 0, both LAC and frequency assign small weight values while for SVM and RELIEF the weight values are set to 0. On the contrary, the weight values of 36 and 166 are set to 0 for LAC and frequency but are not set to 0 in SVM and RELIEF. The correlation of LAC and frequency can be explained by the principle of link-based weighting utual reinforcement. As expected, the rank and weight of features in the LAC and frequency are different. In Table 5, all features are ordered by ascending weight. 69 features (bold) are promoted and4. Rule InterpretationOur recently submitted paper [48] showed that the rules generated by associative classification based on chemical fingerprints and properties can be interpreted by chemical knowledge and shed a light on the molecule design. In this study, we focus on the analysis of association rules generated by LAC using the bio fingerprint (NCI-60 dataset). The analysis for those generated by frequency can be done in the same manner. The accuracy of both frequency and LAC are 99.93 (Table 6) and the average size of the classifier is around 350 rules. For all ten models, the top 5 rules are the same but with different order, support and confidence. The intuitive explanation of Rule 1 in Table 8 is that if compound is inactive to MCF7 andMining by Link-Based Associative Classifier (LAC)HL60 (TB) then it will be inactive to T47D at the same time. The adjusted weighted support of this rule is 29.1 and weighted confidence is 95.9 . Among the 5,937 compounds, 1730 compounds are covered by this rule. All these cell lines in the top 5 rules fall into two categories: a) breast cancer and b) Leukemia. On one hand, it means that there are many compounds which are inactive neither to breast cancer cell lines nor to Leukemia cell lines; on the other hand, it suggests that there might be some associations between these two types of cancers. [49,50] clustered the cell lines based on their gene expression data, their results also indicated that the cell lines in these two categories were clustered into one or their Pentagastrin web clusters were very close to each other. The association of MCF7 and T47D is not surprising as they belong to the same category reast cancer. The rules here may also provide a potential direction of the drug 520-26-3 chemical information resistance of breast cancer and leukemia. [50?2] discovered a novel ABC transporter, breast cancer resistance protein (BCRP). This transporter was termed breast cancer resistance protein (BCRP) because of its identification in MCF-7 human breast carcinoma cells. The drugsensitive cells become drug-resistant cells after transfection or overexpression of BCRP. They also found that 1379592 relatively high expression of BCRP mRNA were observed in around 30 acute myeloid leukemia (AML) cases and suggested a novel mechanism of drug resistance in leukemia. A hybrid feature set integrating the chemical fingerprint and bio fingerprint is generated by combining the MDL public keys and the bio fingerprint. Since we are only interested in the compounds which are active against tumor cell lines, the “inactive” value of the bioassay is treated as a feature of “not existed” in.Ated as the coefficient is 0.949 and 0.958 respectively. The coefficients of SVM, RELIEF and LAC with frequency are greater than 0.75 indicating that all are correlated with frequency. Among them, LAC has the strongest correlation (0.947) with frequency. This is mainly caused by bit 3, 8, 11, 36 and 166. For bit 3, 8 and 11, since their frequencies are not 0, both LAC and frequency assign small weight values while for SVM and RELIEF the weight values are set to 0. On the contrary, the weight values of 36 and 166 are set to 0 for LAC and frequency but are not set to 0 in SVM and RELIEF. The correlation of LAC and frequency can be explained by the principle of link-based weighting utual reinforcement. As expected, the rank and weight of features in the LAC and frequency are different. In Table 5, all features are ordered by ascending weight. 69 features (bold) are promoted and4. Rule InterpretationOur recently submitted paper [48] showed that the rules generated by associative classification based on chemical fingerprints and properties can be interpreted by chemical knowledge and shed a light on the molecule design. In this study, we focus on the analysis of association rules generated by LAC using the bio fingerprint (NCI-60 dataset). The analysis for those generated by frequency can be done in the same manner. The accuracy of both frequency and LAC are 99.93 (Table 6) and the average size of the classifier is around 350 rules. For all ten models, the top 5 rules are the same but with different order, support and confidence. The intuitive explanation of Rule 1 in Table 8 is that if compound is inactive to MCF7 andMining by Link-Based Associative Classifier (LAC)HL60 (TB) then it will be inactive to T47D at the same time. The adjusted weighted support of this rule is 29.1 and weighted confidence is 95.9 . Among the 5,937 compounds, 1730 compounds are covered by this rule. All these cell lines in the top 5 rules fall into two categories: a) breast cancer and b) Leukemia. On one hand, it means that there are many compounds which are inactive neither to breast cancer cell lines nor to Leukemia cell lines; on the other hand, it suggests that there might be some associations between these two types of cancers. [49,50] clustered the cell lines based on their gene expression data, their results also indicated that the cell lines in these two categories were clustered into one or their clusters were very close to each other. The association of MCF7 and T47D is not surprising as they belong to the same category reast cancer. The rules here may also provide a potential direction of the drug resistance of breast cancer and leukemia. [50?2] discovered a novel ABC transporter, breast cancer resistance protein (BCRP). This transporter was termed breast cancer resistance protein (BCRP) because of its identification in MCF-7 human breast carcinoma cells. The drugsensitive cells become drug-resistant cells after transfection or overexpression of BCRP. They also found that 1379592 relatively high expression of BCRP mRNA were observed in around 30 acute myeloid leukemia (AML) cases and suggested a novel mechanism of drug resistance in leukemia. A hybrid feature set integrating the chemical fingerprint and bio fingerprint is generated by combining the MDL public keys and the bio fingerprint. Since we are only interested in the compounds which are active against tumor cell lines, the “inactive” value of the bioassay is treated as a feature of “not existed” in.